Update gpurt from commit 34a59e34

[Continuations] Fix debug traversal path Convert HwPipePoint to PipelineStageFlag Update Trace Source interface [Continuations] Use packed 32-bit shader IDs [Continuations] BoxHeuristicMode changes in one _cont_TraceRay call Consistently use EnqueueDead for dead lanes [Continuations] Add RayHistory data for debug traversal path [Continuations] Conditionally pass 32-bit addresses to the middle-end. [Continuations] Fix debug traversal path compilation Reduce GL2 flushes in AS-related barriers Update Pal::CmdDispatch to PAL ver 909 Fix decodeAS with multiple procedural geometry descs Fix Batch Builder when compiling to SPIRV Use GetRayId() to handle rayquery in graphics shader Make Intrinsics more robust [llvmraytracing] Pass dummy return address to AwaitTraversal [Continuations] Remove !PASS_32_BIT_CR cases Simplify dead lanes workflow, cleanup Traversal Fix for Hitman3 hang on Navi3 [Continuations] Do not set priority in LLPC path Changed minimum scratch buffer size to 8 bytes [Continuations] Simplify Vpc32::isValid()
GPUOpen-Drivers · Dec 4, 2024 · aa5ba68 · aa5ba68
1 parent f2d96f1
commit aa5ba68
Show file tree

Hide file tree

Showing 31 changed files with 1,096 additions and 892 deletions.
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -99,6 +99,11 @@ option(GPURT_BUILD_CONTINUATION "GpuRt uses continuation traversal" ON)
 if (GPURT_BUILD_CONTINUATION)
     gpurt_add_compile_definitions(GPURT_BUILD_CONTINUATION=1)
 endif()
+
+cmake_dependent_option(GPURT_DEBUG_CONTINUATION_TRAVERSAL "Debug continuation traversal on legacy indirect path" OFF "GPURT_BUILD_CONTINUATION" OFF)
+if (GPURT_DEBUG_CONTINUATION_TRAVERSAL)
+    gpurt_add_compile_definitions(GPURT_DEBUG_CONTINUATION_TRAVERSAL=1)
+endif()
 #endif
 
 # Disable run time type information

diff --git a/backends/pal/gpurtPalBackend.cpp b/backends/pal/gpurtPalBackend.cpp
@@ -28,25 +28,6 @@
 namespace GpuRt
 {
 
-// =====================================================================================================================
-// GPURT to PAL enum conversions without undefined behavior.
-static Pal::HwPipePoint GpuRtToPalHwPipePoint(
-    HwPipePoint gpurtHwPipePoint)
-{
-#define HWPIPEPOINTCASE(x) case static_cast<uint32>(Pal::HwPipePoint::x): return Pal::HwPipePoint::x
-    switch (static_cast<uint32>(gpurtHwPipePoint))
-    {
-        HWPIPEPOINTCASE(HwPipeTop);
-        HWPIPEPOINTCASE(HwPipePreCs);
-        HWPIPEPOINTCASE(HwPipeBottom);
-        default:
-            PAL_ASSERT_ALWAYS_MSG("Unhandled HwPipePoint value in conversion: %u\n",
-                                  static_cast<uint32>(gpurtHwPipePoint));
-            return Pal::HwPipePoint::HwPipeTop;
-    }
-#undef HWPIPEPOINTCASE
-}
-
 // =====================================================================================================================
 static Pal::ImmediateDataWidth GpuRtToPalImmediateDataWidth(
     ImmediateDataWidth gpurtImmediateDataWidth)
@@ -132,7 +113,11 @@ void PalBackend::Dispatch(
     uint32                z
     ) const
 {
+#if PAL_INTERFACE_MAJOR_VERSION >= 909
+    GetCmdBuffer(cmdBuffer)->CmdDispatch({ x, y, z }, {});
+#else
     GetCmdBuffer(cmdBuffer)->CmdDispatch({ x, y, z });
+#endif
 }
 
 // =====================================================================================================================
@@ -238,6 +223,7 @@ void PalBackend::InsertBarrier(
 {
     const bool syncDispatch     = flags & BarrierFlagSyncDispatch;
     const bool syncIndirectArgs = flags & BarrierFlagSyncIndirectArg;
+    const bool syncPreCpWrite   = flags & BarrierFlagSyncPreCpWrite;
     const bool syncPostCpWrite  = flags & BarrierFlagSyncPostCpWrite;
 
     Pal::ICmdBuffer* pCmdBuffer = GetCmdBuffer(cmdBuffer);
@@ -247,8 +233,16 @@ void PalBackend::InsertBarrier(
 
     if (syncDispatch || syncIndirectArgs)
     {
-        memoryBarrier.srcStageMask  = Pal::PipelineStageCs;
-        memoryBarrier.srcAccessMask = Pal::CoherShader;
+        memoryBarrier.srcStageMask  |= Pal::PipelineStageCs;
+        memoryBarrier.srcAccessMask |= Pal::CoherShader;
+    }
+
+    if (syncPreCpWrite)
+    {
+        memoryBarrier.srcStageMask  |= Pal::PipelineStagePostPrefetch;
+        memoryBarrier.srcAccessMask |= Pal::CoherShader;
+        memoryBarrier.dstStageMask  |= Pal::PipelineStagePostPrefetch;
+        memoryBarrier.dstAccessMask |= Pal::CoherCp;
     }
 
     if (syncPostCpWrite)
@@ -359,12 +353,11 @@ void PalBackend::UpdateMemory(
 // =====================================================================================================================
 void PalBackend::WriteTimestamp(
     ClientCmdBufferHandle  cmdBuffer,
-    HwPipePoint            hwPipePoint,
     const Pal::IGpuMemory& timeStampVidMem,
     uint64                 offset
     ) const
 {
-    GetCmdBuffer(cmdBuffer)->CmdWriteTimestamp(GpuRtToPalHwPipePoint(hwPipePoint), timeStampVidMem, offset);
+    GetCmdBuffer(cmdBuffer)->CmdWriteTimestamp(Pal::PipelineStageBottomOfPipe, timeStampVidMem, offset);
 }
 
 // =====================================================================================================================

diff --git a/backends/pal/gpurtPalBackend.h b/backends/pal/gpurtPalBackend.h
@@ -121,7 +121,6 @@ class PalBackend : public IBackend
 
     virtual void WriteTimestamp(
         ClientCmdBufferHandle  cmdBuffer,
-        HwPipePoint            hwPipePoint,
         const Pal::IGpuMemory& timeStampVidMem,
         uint64                 offset
     ) const override;

diff --git a/gpurt/gpurt.h b/gpurt/gpurt.h
@@ -737,7 +737,7 @@ struct DeviceSettings
     uint32                      numRebraidIterations;
     uint32                      rebraidQualityHeuristic;
 
-    uint32                      plocRadius;                           // PLOC Radius
+    uint32                      plocRadius;                           // PLOC nearest neighbor search adius
     uint32                      maxTopDownBuildInstances;             // Max instances allowed for top down build
     uint32                      parallelBuildWavesPerSimd;            // Waves per SIMD to launch for parallel build
 

diff --git a/gpurt/gpurtBackend.h b/gpurt/gpurtBackend.h
@@ -75,21 +75,13 @@ struct BufferViewInfo
     BufferViewSwizzle swizzle;
 };
 
-// =====================================================================================================================
-// Copy of Pal::HwPipePoint with values we use.
-enum class HwPipePoint : uint32
-{
-    HwPipeTop    = 0x0,
-    HwPipePreCs  = 0x1,
-    HwPipeBottom = 0x7,
-};
-
 // =====================================================================================================================
 enum BarrierFlags : uint32
 {
     BarrierFlagSyncDispatch    = 0x1, // Stall the following dispatch until all previous dispatch done
     BarrierFlagSyncIndirectArg = 0x2, // Prepare previous shader output for indirect argument use
-    BarrierFlagSyncPostCpWrite = 0x4, // Prepare data set by CP for shader use
+    BarrierFlagSyncPreCpWrite  = 0x4, // Prepare for CP write
+    BarrierFlagSyncPostCpWrite = 0x8, // Prepare data set by CP for shader use
 };
 
 // =====================================================================================================================
@@ -185,7 +177,6 @@ class IBackend
     // Will eventually replaced with a callback or other abstraction to avoid referencing video memory.
     virtual void WriteTimestamp(
         ClientCmdBufferHandle  cmdBuffer,
-        HwPipePoint            hwPipePoint,
         const Pal::IGpuMemory& timeStampVidMem,
         uint64                 offset) const = 0;
 

diff --git a/gpurt/gpurtBuildSettings.h b/gpurt/gpurtBuildSettings.h
@@ -62,7 +62,7 @@ struct CompileTimeBuildSettings
     uint32 radixSortScanLevel;
     uint32 emitCompactSize;
     uint32 enableBVHBuildDebugCounters;
-    uint32 plocRadius;
+    uint32 nnSearchRadius;
     uint32 enablePairCostCheck;
     uint32 enableVariableBitsMortonCode;
     uint32 rebraidType;
@@ -112,7 +112,7 @@ struct CompileTimeBuildSettings
 #define BUILD_SETTINGS_DATA_RADIX_SORT_SCAN_LEVEL_ID                  7
 #define BUILD_SETTINGS_DATA_EMIT_COMPACT_SIZE_ID                      8
 #define BUILD_SETTINGS_DATA_ENABLE_BVH_BUILD_DEBUG_COUNTERS_ID        9
-#define BUILD_SETTINGS_DATA_PLOC_RADIUS_ID                            10
+#define BUILD_SETTINGS_DATA_NN_SEARCH_RADIUS_ID                       10
 #define BUILD_SETTINGS_DATA_ENABLE_PAIR_COST_CHECK_ID                 11
 #define BUILD_SETTINGS_DATA_ENABLE_VARIABLE_BITS_MC_ID                12
 #define BUILD_SETTINGS_DATA_REBRAID_TYPE_ID                           13

diff --git a/gpurt/gpurtInlineFuncs.h b/gpurt/gpurtInlineFuncs.h
@@ -156,23 +156,6 @@ inline BufferViewFormat GetSingleComponentFormatForFormat(BufferViewFormat forma
     }
 }
 
-//=====================================================================================================================
-// Converts the value of a Pal::HwPipePoint into a GpuRt::HwPipePoint without undefined behavior.
-inline HwPipePoint PalToGpuRtHwPipePoint(uint32 palHwPipePoint)
-{
-#define HWPIPEPOINTCASE(x) case static_cast<uint32>(HwPipePoint::x): return HwPipePoint::x
-    switch (palHwPipePoint)
-    {
-        HWPIPEPOINTCASE(HwPipeTop);
-        HWPIPEPOINTCASE(HwPipePreCs);
-        HWPIPEPOINTCASE(HwPipeBottom);
-        default:
-            PAL_ASSERT_ALWAYS_MSG("Unhandled HwPipePoint value in conversion: %u\n", palHwPipePoint);
-            return HwPipePoint::HwPipeTop;
-    }
-#undef HWPIPEPOINTCASE
-}
-
 //=====================================================================================================================
 // Return the number of components for a buffer view format when it's used as a vertex format.
 inline uint8 GetNumComponentsForVertexFormat(VertexFormat format)

diff --git a/src/gpurtBvhBatcher.cpp b/src/gpurtBvhBatcher.cpp
@@ -104,7 +104,7 @@ void BvhBatcher::BuildAccelerationStructureBatch(
             // but otherwise do not participate in the rest of the build.
             if (isUpdate)
             {
-                builder.EmitPostBuildInfo();
+                builder.EmitPostBuildInfoDispatch();
             }
             else
             {
@@ -146,7 +146,11 @@ void BvhBatcher::BuildAccelerationStructureBatch(
     {
         RGP_PUSH_MARKER("Process Empty BVH builds");
         DispatchInitAccelerationStructure<false>(emptyBuilders);
-        BuildPhase(emptyBuilders, &BvhBuilder::EmitPostBuildInfo);
+        if (PhaseEnabled(BuildPhaseFlags::SeparateEmitPostBuildInfoPass))
+        {
+            Barrier();
+            BuildPhase(emptyBuilders, &BvhBuilder::EmitPostBuildInfoDispatch);
+        }
         RGP_POP_MARKER();
     }
 
@@ -264,17 +268,10 @@ void BvhBatcher::BuildRaytracingAccelerationStructureBatch(
     {
         RGP_PUSH_MARKER("EmitPostBuildInfo");
         Barrier();
-        BuildPhase("Updates", updaters, &BvhBuilder::EmitPostBuildInfo);
-        BuildPhase("Builds", builders, &BvhBuilder::EmitPostBuildInfo);
-
+        BuildPhase(BuildPhaseFlags::SeparateEmitPostBuildInfoPass, updaters, &BvhBuilder::EmitPostBuildInfoDispatch);
+        BuildPhase(BuildPhaseFlags::SeparateEmitPostBuildInfoPass, builders, &BvhBuilder::EmitPostBuildInfoDispatch);
         RGP_POP_MARKER();
     }
-    else
-    {
-        // Execute EmitPostBuildInfo without any RGP markers
-        BuildPhase(updaters, &BvhBuilder::EmitPostBuildInfo);
-        BuildPhase(builders, &BvhBuilder::EmitPostBuildInfo);
-    }
 
     if (PhaseEnabled(BuildPhaseFlags::BuildDumpEvents))
     {